LKC-Net: large kernel convolution object detection network.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Current deep learning object detection methods face challenges due to small kernel convolutions limiting receptive fields.
• This limitation hinders semantic feature extraction, leading to issues like missed or incorrect detections.

Purpose of the Study:

• To propose LKC-Net, a large kernel convolution object detection network designed to enhance feature capture and receptive fields.
• To address the limitations of small kernel convolutions in improving object detection accuracy.

Main Methods:

• Introduced a feature capture enhancement block utilizing large kernel convolution and depth convolution for improved semantic feature extraction and parameter efficiency.
• Developed a vast receptive field attention mechanism to boost channel-wise information extraction, demonstrating superior compatibility with the proposed backbone.
• Enhanced the loss function with SIoU to resolve angle mismatches between ground truth and predicted bounding boxes.

Main Results:

• LKC-Net demonstrated improved semantic feature capturing ability compared to existing methods.
• The vast receptive field attention mechanism showed enhanced channel information extraction.
• Experiments on Pascal VOC and MS COCO datasets validated the effectiveness of LKC-Net.

Conclusions:

• LKC-Net effectively overcomes the limitations of small kernel convolutions in object detection.
• The proposed network achieves superior performance by enhancing feature capture and utilizing an effective attention mechanism.
• The integration of SIoU loss further refines detection accuracy.